A method and system for identifying low-altitude flying object threats based on dynamic density clustering

Abstract

This invention proposes a method and system for identifying low-altitude flying objects (LAO) threats based on dynamic density clustering, relating to the field of rail transit safety technology. Addressing the problems of blind spots in single-sensor monitoring, poor adaptability of fixed-parameter clustering algorithms to dynamic targets, and the lack of quantitative assessment and graded response mechanisms for threat levels in existing technologies, this invention fuses millimeter-wave radar and infrared thermal imaging data to form a unified spatiotemporal point cloud. It employs an improved clustering algorithm based on velocity and density adaptive determination of cluster numbers to achieve accurate target segmentation, and combines temperature and morphological characteristics for LAO classification, constructing a multi-factor threat model to achieve dynamic threat level calculation and graded early warning. This invention improves the accuracy and precision of LAO identification and enhances robustness in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of rail transit safety technology, and in particular to a method and system for identifying low-altitude flying object threats based on dynamic density clustering. Background Technology

[0002] With the rapid expansion of urban rail transit networks and the continuous increase in operational density, open areas such as elevated subway sections and tunnel entrances are facing an increasing risk of interference from low-altitude flying objects, such as flocks of birds, stray or unauthorized drone incursions. Real-time and accurate identification and threat assessment of low-altitude flying objects is of significant importance for preventing collisions, ensuring train safety, and improving operational efficiency, and has become an important development direction in the field of rail transit safety monitoring.

[0003] However, existing low-altitude flying object monitoring technologies still have significant limitations. On the one hand, at the perception level, they largely rely on single-type sensors. Visual methods experience a sharp decline in performance at night, in low light conditions such as rain or fog, or in adverse weather conditions. Traditional radar struggles to distinguish between biological and non-biological targets and is susceptible to environmental clutter, leading to both blind spots and false alarms. On the other hand, at the data processing level, conventional target detection and clustering algorithms often use fixed parameters, making it difficult to adapt to the dynamic characteristics of targets such as high-speed drones and dense flocks of birds. This results in insufficient accuracy and robustness in target segmentation, easily leading to false detections and missed detections.

[0004] Therefore, there is an urgent need for a low-altitude flying object threat identification method based on dynamic density clustering with higher identification accuracy and stronger robustness. Summary of the Invention

[0005] To address this, the present invention provides a method and system for identifying low-altitude flying object threats based on dynamic density clustering, which overcomes the shortcomings of existing technologies in identifying threats under low light conditions such as nighttime, rain, and fog, or in adverse weather conditions; traditional radar has difficulty distinguishing between biological and non-biological targets and is easily affected by environmental clutter, resulting in both monitoring blind spots and false alarms; it is also difficult to adapt to the dynamic characteristics of targets such as high-speed drones and dense flocks of birds, and its accuracy and robustness in target segmentation are insufficient, easily leading to false detections and missed detections.

[0006] To achieve the above objectives, this invention provides a method for identifying low-altitude flying object threats based on dynamic density clustering, comprising: S1, Install several monitoring devices at preset intervals in the target area to obtain multimodal monitoring data of the target area; S2, preprocess the multimodal monitoring data, and fuse the preprocessed multimodal monitoring data to generate fused point cloud data; S3, The fused point cloud data is clustered and segmented based on an improved clustering algorithm to generate independent clusters; S4. Obtain the temperature and morphological features of each independent cluster, and determine the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data. S5. Based on the type of the flying object, combined with the position and speed of the flying object, calculate the threat level of the flying object; S6. Determine the risk level based on the threat level of the flying object, and execute the corresponding response strategy based on the risk level.

[0007] Furthermore, the installation of several monitoring devices at preset intervals in the target area includes: S11, At the elevated section of the subway and / or the entrance of the tunnel in the target area, monitoring nodes consisting of millimeter-wave radar and infrared thermal imagers are deployed at preset intervals to form an electronic space curtain covering the target area. S12, radar point cloud data and infrared thermal image data are synchronously collected through the monitoring node as the multimodal monitoring data; S13, based on oblique photography technology, collects real-world images of the surrounding area of ​​the target region and builds them on a pre-set digital twin platform; S14. Based on BIM technology, model the subway vehicles and subway infrastructure to generate a BIM model, and build the BIM model on a preset digital twin platform.

[0008] Furthermore, the preprocessing of the multimodal monitoring data and the fusion of the preprocessed multimodal monitoring data to generate fused point cloud data include: S21, preprocess the multimodal monitoring data. The preprocessing includes data cleaning and spatiotemporal alignment. Data cleaning is used to remove invalid and redundant data. Spatiotemporal alignment uses the extended Kalman filter method to unify the monitoring data from different monitoring devices into the same spatiotemporal coordinate system. S22 fuses the preprocessed multimodal monitoring data to generate fused point cloud data containing three-dimensional spatial coordinates, velocity, and temperature information.

[0009] Furthermore, the improved clustering algorithm is used to cluster and segment the fused point cloud data to generate independent clusters, including: S31, Calculate the velocity distribution variance based on the velocity of each point cloud data in the fused point cloud data; S32, Calculate the density value of the point cloud data based on the effective number of points and the detection area; wherein, the effective number of points is the sum of the number of point cloud data with a speed greater than a preset speed threshold; S33, Based on the velocity distribution variance and the density value of the point cloud data, determine the number of clusters for clustering; S34. Based on the cluster centroid positions of historical frames, calculate the minimum cumulative distance sum of each candidate point in the current frame, and select the initial centroid based on the minimum cumulative distance sum; S35, Construct a loss function that integrates spatiotemporal constraints, the loss function including a spatial distance term, a velocity consistency term, and a volume rationality term; S36, with the goal of minimizing the loss function, assigns each point cloud data to the cluster that minimizes its own loss distance, thus completing the clustering assignment; S37, based on the allocation results, update the spatial centroid position, average velocity, and point cloud convex hull volume of each cluster; S38, calculate the current total loss value, and determine whether the difference between the current total loss value and the total loss value of the previous iteration is less than a preset threshold: if yes, stop the iteration and output the independent cluster; if no, return to step S36 to continue the iteration.

[0010] Furthermore, the step of acquiring the temperature and morphological features of each independent cluster, and determining the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data, includes: S41, extract the temperature features of each independent cluster, the temperature features including the average temperature of the independent cluster and the maximum temperature difference within the independent cluster; S42, extract the morphological features of each independent cluster, the morphological features including the volume and surface curvature variance calculated based on the point cloud convex hull of the independent cluster; S43, based on the temperature and morphological characteristics, determine the type of the flying object according to a preset classification rule; wherein, the preset classification rule is that if the average temperature is lower than a first temperature threshold and the surface curvature variance is higher than a first curvature threshold, then the flying object is determined to be a bird; if the maximum temperature difference is greater than a second temperature threshold and the surface curvature variance is lower than a second curvature threshold, then the flying object is determined to be a drone. S44, Calculate the confidence level of the classification decision based on the temperature feature and the morphological feature.

[0011] Furthermore, if the confidence level of the classification decision is lower than a preset confidence threshold, an adaptive adjustment to the independent clustering is triggered, the adaptive adjustment including: If the temperature gradient within the independent cluster is greater than the preset gradient threshold, then the point cloud data of the independent cluster is clustered a second time according to the temperature gradient to generate a new independent cluster. If the volume of the independent cluster exceeds the preset reasonable volume threshold, the number of clusters determined in S3 is increased, and S33 to S38 are re-executed based on the increased number of clusters to generate a new independent cluster. Based on the new independent clusters obtained, S41 to S43 are re-executed.

[0012] Furthermore, the calculation of the threat level of the flying object based on its type, position, and speed includes: S51, acquire the position information and velocity vector of the flying object, and calculate the minimum spatial distance between the flying object and the center line of the orbit based on the position of the flying object; S52, Based on the velocity vector, calculate the tangential velocity component of the flying object in the direction approaching the orbit; S53, determine the corresponding type weighting coefficient according to the type of the flying object; S54, the threat level of the flying object is calculated by weighted summation based on the minimum spatial distance, the tangential velocity component, and the type weight coefficient; wherein, the weighted summation calculation process is to multiply the reciprocal of the minimum spatial distance term, the tangential velocity component term, and the type weight term after normalization by their respective preset weight coefficients and then sum them.

[0013] Furthermore, the risk level determination based on the threat level of the flying object includes: If the threat level of the flying object is greater than the first threat threshold, it is determined to be of a high-risk level; If the threat level of the flying object is greater than or equal to the second threat threshold and less than or equal to the first threat threshold, it is determined to be of medium risk level; If the threat level of the flying object is less than the second threat threshold, it is determined to be of low risk level.

[0014] Furthermore, the response strategy includes performing trajectory backtracking and emergency plan simulation on the preset digital twin platform.

[0015] This invention also provides a low-altitude flying object threat identification system based on dynamic density clustering, the system being used to implement any of the low-altitude flying object threat identification methods based on dynamic density clustering, the system comprising: The information acquisition module is used to install several monitoring devices at preset intervals in the target area to obtain multimodal monitoring data of the target area. The data fusion module, connected to the information acquisition module, is used to preprocess the multimodal monitoring data and fuse the preprocessed multimodal monitoring data to generate fused point cloud data. The clustering and segmentation module, connected to the data fusion module, is used to perform clustering and segmentation on the fused point cloud data based on an improved clustering algorithm to generate independent clusters; The clustering identification module, connected to the clustering segmentation module, is used to acquire the temperature and morphological features of each independent cluster, and determine the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data. The threat calculation module, connected to the clustering and identification module, is used to calculate the threat level of the flying object based on its type, position, and speed. The risk assessment module, connected to the threat calculation module, is used to determine the risk level based on the threat level of the flying object and to execute corresponding response strategies based on the risk level.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: Firstly, this invention addresses the bottleneck of existing clustering algorithms that use fixed parameters and struggle to adapt to the dynamic characteristics of targets. It proposes an improved clustering algorithm that dynamically determines the number of clusters based on velocity distribution variance and spatial density, and constructs a spatiotemporal constraint loss function that integrates spatial distance, velocity consistency, and volume rationality. This algorithm can adapt to the motion and distribution characteristics of targets such as high-speed UAVs and dense flocks of birds, significantly improving the accuracy and robustness of clustering and segmentation, and effectively reducing false detection and false negative rates caused by noise interference or target adhesion.

[0017] Secondly, this invention constructs an intelligent feedback loop by introducing a clustering secondary optimization mechanism triggered by classification confidence. Specifically, when the automatic classification confidence is insufficient, the system can automatically diagnose the root cause of the problem; if the temperature gradient within a cluster is abnormal, secondary clustering based on the temperature gradient is initiated to separate regions with different thermal characteristics within the target, improving the recognition accuracy of complex targets; if the cluster volume exceeds a reasonable threshold, the number of clusters is dynamically increased and the clusters are re-segmented to effectively separate densely clustered targets, enhancing individual identification capabilities. This mechanism transforms the traditional one-time, fixed processing flow into an adaptive, continuously optimized intelligent process, significantly reducing reliance on manual review and improving the automated processing capabilities for fuzzy samples, complex targets, and extreme scenarios, as well as robustness in complex environments.

[0018] Thirdly, the overall solution provided by this invention deeply integrates multi-source sensing, dynamic intelligent processing, and digital twin decision-making. At the sensing level, through the coordinated deployment and data fusion of millimeter-wave radar and infrared thermal imager, it effectively overcomes the performance limitations of single sensors under complex weather and lighting conditions, achieving all-weather, highly reliable target detection. At the processing level, the innovatively proposed dynamic adaptive clustering algorithm and the secondary optimization mechanism based on confidence feedback solve the problem of poor adaptability of traditional methods to high-speed, dense, and dynamic target segmentation, significantly improving the accuracy and robustness of target separation and feature extraction. At the decision-making level, by constructing a multi-factor threat model that integrates target type, location, and speed, and combining it with a digital twin platform to achieve visual mapping and system linkage, this invention improves the real-time identification accuracy, threat assessment scientificity, and emergency response automation level of low-altitude flying object intrusions in subway protected areas, providing efficient and reliable technical support for ensuring the safe operation of urban rail transit. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a method for identifying low-altitude flying object threats based on dynamic density clustering, provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a low-altitude flying object threat identification system based on dynamic density clustering, provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0021] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0022] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0023] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] Example 1 like Figure 1 As shown, this invention provides a method for identifying low-altitude flying object threats based on dynamic density clustering, comprising: S1, Install several monitoring devices at preset intervals in the target area to obtain multimodal monitoring data of the target area; The installation of several monitoring devices at preset intervals in the target area includes: S11, At the elevated section of the subway and / or the entrance of the tunnel in the target area, monitoring nodes consisting of millimeter-wave radar and infrared thermal imagers are deployed at preset intervals to form an electronic space curtain covering the target area. S12, radar point cloud data and infrared thermal image data are synchronously collected through the monitoring node as the multimodal monitoring data; S13, based on oblique photography technology, collects real-world images of the surrounding area of ​​the target region and builds them on a pre-set digital twin platform; S14. Based on BIM technology, model the subway vehicles and subway infrastructure to generate a BIM model, and build the BIM model on a preset digital twin platform.

[0025] In one possible implementation, integrated monitoring nodes are fixedly installed at preset intervals of 50 to 100 meters in key sections of the target area, such as both sides of elevated subway sections and above tunnel entrances. Each monitoring node includes at least one millimeter-wave radar, preferably in the 77 GHz band, with speed and micro-motion detection capabilities, and one uncooled infrared thermal imager. The two are structurally integrated and jointly calibrated to ensure complementary field of view and detection range. These nodes are interconnected through wired or wireless communication networks, forming a continuous, seamless electronic space wall above the physical protection zone, achieving comprehensive coverage and collaborative perception of the low-altitude area.

[0026] The system controls each monitoring node to acquire data synchronously or quasi-synchronously. The millimeter-wave radar outputs a point cloud sequence in real time, containing target range, azimuth, elevation, radial velocity, and reflection intensity; the infrared thermal imager outputs a sequence of infrared thermal images with timestamps, where each pixel or region contains temperature information. Both types of data streams are uploaded in real time to edge computing nodes or a central server via a built-in communication module, serving as raw multimodal monitoring data for subsequent processing.

[0027] A drone equipped with a five-lens oblique photography camera was used to conduct aerial photography of the protected area along the subway line and its surrounding environment, collecting multi-view, high-overlap image data. Photogrammetric processing software was used to perform aerial triangulation and dense matching on the images, generating a high-precision real-scene 3D mesh model. This model was imported and embedded into the system's dedicated digital twin platform, serving as the real geographical background for visualization and spatial analysis.

[0028] It should be noted that digital twin platforms are a conventional technical means familiar to those skilled in the art; therefore, the process of building a digital twin platform will not be described in detail here.

[0029] Using Building Information Modeling (BIM) technology, a detailed BIM model is created based on the subway line's design drawings and asset data. This model includes tracks, overhead contact lines, sound barriers, station structures, and operating vehicles. The model contains geometric information and semantic attributes, such as equipment type, ID, and maintenance status. This BIM model is then imported into the aforementioned digital twin platform and precisely aligned and merged with the constructed real-world 3D mesh model within a unified coordinate system. This recreates a geometrically and semantically complete virtual subway environment in digital space, providing a foundation for subsequent threat visualization, trajectory mapping, and situational awareness coordination.

[0030] This invention constructs an electronic spatial curtain by deploying monitoring nodes integrating millimeter-wave radar and infrared thermal imagers at preset intervals in key areas such as subway elevated sections and tunnel entrances. This solves the monitoring blind spots and limitations of single sensors in complex environments, improving the all-weather, all-time monitoring coverage and system reliability for low-altitude flying objects, and reducing the false alarm rate caused by the single performance of sensors. By simultaneously collecting radar point cloud data and infrared thermal image data as multimodal monitoring data, it solves the problem of insufficient information dimensions and inability to support subsequent fine classification and threat assessment by traditional single data sources, improving the information richness and spatiotemporal consistency of the data, and laying a data foundation for subsequent accurate target identification and behavior analysis. By using oblique photography technology to create a realistic 3D model of the target area and integrating it into a digital twin platform, the problem of traditional 2D monitoring images lacking depth information and having unintuitive spatial relationships is solved. This improves the realism of the monitoring scene and the intuitive understanding of the spatial situation, reducing the difficulty for maintenance personnel to interpret abstract data or 2D drawings. By merging and aligning the realistic 3D model with the BIM model in the digital twin platform, the problem of separating the geographical background and facility model and making it difficult to conduct integrated spatial analysis and display is solved. This improves the integrity and realism of the digital twin environment, laying a unified spatial foundation for realizing intelligent security and operation and maintenance management with virtual-real mapping and real-time interaction, and reducing the complexity of data fusion and collaboration between multiple systems.

[0031] S2, preprocess the multimodal monitoring data, and fuse the preprocessed multimodal monitoring data to generate fused point cloud data; The preprocessing of the multimodal monitoring data and the fusion of the preprocessed multimodal monitoring data include: S21, preprocess the multimodal monitoring data. The preprocessing includes data cleaning and spatiotemporal alignment. Data cleaning is used to remove invalid and redundant data. Spatiotemporal alignment uses the extended Kalman filter method to unify the monitoring data from different monitoring devices into the same spatiotemporal coordinate system. S22 fuses the preprocessed multimodal monitoring data to generate spatiotemporal fused point cloud data containing three-dimensional spatial coordinates, velocity, and temperature information.

[0032] In one possible implementation, the received raw radar point cloud data and infrared thermal image data undergo preliminary screening and purification. For the radar point cloud data, invalid noise points caused by ground clutter, rain, fog, insects, etc., are removed by setting reflection intensity thresholds, range thresholds, and Doppler velocity filters, and obviously stationary background points, such as building reflections, are also filtered out. For the infrared thermal image data, background subtraction and temperature threshold filtering are used to suppress redundant thermal noise caused by environmental thermal radiation and to remove instantaneous high-temperature artifacts caused by solar reflection, etc. This step aims to improve data quality and reduce interference in subsequent calculations.

[0033] To achieve accurate fusion of millimeter-wave radar and infrared thermal imager data, they must be unified to the same spatiotemporal reference system. For time alignment, the system assigns high-precision synchronization timestamps to all data, such as using the PTP protocol, to ensure data stream synchronization. For spatial alignment, an extended Kalman filter method is employed for sensor fusion and coordinate unification. Specifically, the target state after fusion at the previous moment, such as position and velocity, is used as the predicted value. The polar coordinates measured by the radar at the current moment, such as range, azimuth, elevation, and radial velocity, along with the target angle information calculated from the infrared image using a calibration model, are used as the observed values. Through iterative updates using an extended Kalman filter, the optimal estimate of the target's three-dimensional position (x, y, z), velocity vector, and corresponding temperature information t from the infrared sensor is obtained in a unified world coordinate system, such as the plane coordinate system and elevation datum used in railway line design. This solves the problem of temporal and spatial inconsistencies in heterogeneous sensor data.

[0034] After preprocessing and spatiotemporal alignment, the system performs attribute-level fusion of radar point cloud data and infrared temperature information. For each data point after extended Kalman filtering, the system encapsulates it into a structured data unit, called fused point cloud data. Each data point contains the following core attributes: x, y, z: The three-dimensional spatial coordinates of the target in a unified world coordinate system, in meters; v, the radial velocity of the target or the velocity component estimated by EKF fusion, in meters per second; t is the temperature value of the spatial location point obtained by mapping or interpolation from the registered infrared data, in degrees Celsius.

[0035] Finally, a series of fused point cloud data is output, such as Pi=(x, y, z, v, t), forming a spatiotemporally synchronized fused point cloud dataset with rich attributes. This dataset not only contains the precise geometric and motion information of the target, but also incorporates thermal features, providing a multi-dimensional unified data foundation for subsequent dynamic clustering, target classification, and threat assessment.

[0036] This invention cleanses multimodal raw monitoring data, effectively removing invalid noise points and redundant data. This solves the problems of low signal-to-noise ratio and numerous interference factors in raw sensor data, improving the purity and quality of input data and reducing the error rate of subsequent clustering, segmentation, and feature extraction caused by data noise. By employing an extended Kalman filter method for precise spatiotemporal alignment of radar and infrared data, it solves the problem of spatiotemporal inconsistency caused by different installation locations, sampling sequences, and coordinate systems of heterogeneous sensors, achieving millimeter-level / millisecond-level spatiotemporal synchronization accuracy. This improves the reliability and accuracy of multi-source data fusion and provides a unified spatiotemporal reference for subsequent processing. By fusing preprocessed radar point cloud data with infrared temperature information at the attribute level, spatiotemporal fused point cloud data containing multi-dimensional attributes such as three-dimensional coordinates, velocity, and temperature is generated. This solves the problem of insufficient information dimension of single-type data and inability to fully characterize target characteristics, improving the information completeness and representation ability of the data. It provides a key data foundation for subsequent realization of accurate target recognition and classification based on multiple features. By forming a structured fused point cloud dataset, the problems of multi-source data separation and processing and complex correlation in traditional methods are solved. This simplifies the data interface and processing logic of downstream algorithms, improves the efficiency and system integration of the entire processing flow, and reduces the complexity of algorithm development and maintenance.

[0037] S3, The fused point cloud data is clustered and segmented based on an improved clustering algorithm to generate independent clusters; The clustering and segmentation of the fused point cloud data based on the improved clustering algorithm includes: S31, Calculate the velocity distribution variance based on the velocity of each point cloud data in the fused point cloud data; S32, Calculate the density value of the point cloud data based on the effective number of points and the detection area; wherein, the effective number of points is the sum of the number of point cloud data with a speed greater than a preset speed threshold; S33, Based on the velocity distribution variance and the density value of the point cloud data, determine the number of clusters for clustering; S34. Based on the cluster centroid positions of historical frames, calculate the minimum cumulative distance sum of each candidate point in the current frame, and select the initial centroid based on the minimum cumulative distance sum; S35, Construct a loss function that integrates spatiotemporal constraints, the loss function including a spatial distance term, a velocity consistency term, and a volume rationality term; S36, with the goal of minimizing the loss function, assigns each point cloud data to the cluster that minimizes its own loss distance, thus completing the clustering assignment; S37, based on the allocation results, update the spatial centroid position, average velocity, and point cloud convex hull volume of each cluster; S38, calculate the current total loss value, and determine whether the difference between the current total loss value and the total loss value of the previous iteration is less than a preset threshold: if yes, stop the iteration and output the independent cluster; if no, return to step S36 to continue the iteration.

[0038] In one possible implementation, the fused point cloud dataset of the current frame is traversed, the velocity attribute associated with each data point is extracted, and the variance of the velocity of all points in the frame is calculated to quantify the dispersion of the target group's motion. A higher variance usually means that there are multiple targets with different motion patterns in the scene or a single target is performing highly maneuverable motion.

[0039] To assess the aggregation of point clouds within the detection space, the system calculates the spatial density of the point cloud. First, it counts the number of valid points with velocities greater than a preset velocity threshold. Then, based on the radar's theoretical detection range, such as a circular area defined by the maximum detection radius, it calculates the detection area. Finally, the ratio of the number of valid points to the detection area is the density value of the point cloud data, reflecting the density of valid moving points per unit area. A preferred embodiment of the preset velocity threshold is 1 m / s.

[0040] Dynamic decision-making is performed based on the calculation results of steps S31 and S32. When the velocity distribution variance is greater than 5m... 2 / s 2 Then, the number of clusters K in the point cloud data is calculated using the following formula: Where, min represents taking the smaller of the two values, σ is the variance of the velocity distribution, and ρ is the density value of the point cloud data; When the velocity distribution variance is less than 5m 2 / s 2 When the velocity variance is large, i.e., the target motion is complex, or the point cloud density is high, i.e., the target is clustered, the number of clusters K in the point cloud data is increased accordingly to adapt to the possibility of multiple targets or scenarios requiring fine segmentation; at the same time, by taking the smaller value between the calculated value and 5 as the upper limit constraint, too many invalid small clusters are prevented from being generated under noise interference, thus ensuring the efficiency and stability of the algorithm.

[0041] To improve clustering convergence speed and leverage temporal continuity, the system initializes cluster centroid positions by referencing historical frames, such as the previous frame. Let there be H centroids in each historical frame. For each candidate data point in the current frame, calculate its minimum Euclidean distance to all historical centroids. Select the M points that minimize the sum of their minimum distances as initial centroids. If M is greater than H, new centroids are selected from the remaining points, maximizing their minimum distance to the already selected initial centroids to ensure a good spatial distribution of initial centroids.

[0042] A loss function L is defined that comprehensively considers spatial proximity, motion consistency, and physical rationality to guide cluster assignment and centroid updating. This function is expressed as follows: L = Σ (spatial distance term + α × velocity consistency term + β × volume rationality term); Wherein, the spatial distance term is the squared Euclidean distance from the point to the spatial centroid of its cluster K, the velocity consistency term is the absolute difference between the point's velocity and the average velocity of its cluster, and the volume rationality term is defined as max(0,(V k -V max ) / V max ) 2 , where V k Let V be the convex hull volume of the point cloud cluster. max This is a preset maximum reasonable volume threshold for a single target. This penalty applies to clusters with abnormally large volumes to prevent the erroneous merging of multiple neighboring targets. α and β are calibrable weighting coefficients used to balance the importance of the three constraints.

[0043] With the goal of minimizing the total loss function that fuses spatiotemporal constraints, the system traverses all point cloud data and reassigns them to the cluster that minimizes the calculated loss term for each point, completing one round of clustering. Based on the new assignment results, the system updates the key attributes of each cluster, specifically including recalculating the geometric center of all points within the cluster, recalculating the average velocity of all points within the cluster, and calculating the 3D convex hull volume based on the spatial coordinates of the points within the cluster.

[0044] After calculating the current total loss value after steps S36 and S37, determine whether the absolute difference between the current total loss value and the total loss value of the previous iteration is less than a preset convergence threshold. If it is less than the preset convergence threshold, the algorithm is considered to have converged, the iteration stops, and the currently obtained clusters are output as independent clusters, each cluster representing a potential low-altitude flying object target. Otherwise, return to step S36, and start a new round of allocation and update iteration based on the updated cluster attributes until the convergence condition is met.

[0045] This invention addresses the challenge of traditional density-based clustering algorithms, which use fixed neighborhood radii and minimum point counts, failing to adapt to high-speed, highly maneuverable targets like drones and densely clustered targets like flocks of birds, by dynamically determining the number of clusters K for clustering based on the velocity distribution variance and spatial density values ​​of fused point clouds. This improves the algorithm's adaptability to targets with different motion states and distribution patterns, reducing the risk of over-segmentation or under-segmentation due to parameter fixation. Furthermore, by referencing the cluster centroid positions of historical frames for current frame centroid initialization, it solves the problem of slow convergence and unstable results that can occur with random initialization. Utilizing the temporal continuity of target motion improves the rationality of initial cluster positions and the algorithm's convergence speed, enhancing the spatiotemporal consistency of clustering results between adjacent frames and laying the foundation for stable tracking. Finally, by constructing a loss function that integrates spatial distance, velocity consistency, and volume rationality, and using this as the target for iterative clustering, it solves the problem of... Traditional clustering only considers spatial proximity while ignoring the consistency of target motion and physical rationality. This paper achieves optimized segmentation under multi-dimensional constraints of space, time, and physics, significantly improving the segmentation accuracy and robustness for high-speed dynamic targets and densely clustered targets. It effectively reduces the false detection and false negative rates caused by noise interference or target proximity. By introducing the point cloud convex hull volume as a constraint into the loss function and penalizing it, the paper solves the problem of multiple neighboring targets being incorrectly merged into one cluster due to the spatial adhesion of their point clouds. This improves the ability to separate individual targets in dense flying objects and ensures the object independence of subsequent target classification and threat assessment. By setting an iterative convergence condition based on the change of total loss value, the paper solves the problem that a fixed number of iterations may lead to wasted computational resources or insufficient convergence. It achieves adaptive convergence judgment of the algorithm and optimizes the utilization efficiency of computational resources while ensuring the stability of clustering results.

[0046] S4, acquire the temperature and morphological features of each independent cluster, and determine the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data, including: S41, extract the temperature features of each independent cluster, the temperature features including the average temperature of the independent cluster and the maximum temperature difference within the independent cluster; S42, extract the morphological features of each independent cluster, the morphological features including the volume and surface curvature variance calculated based on the point cloud convex hull of the independent cluster; S43, based on the temperature and morphological characteristics, determine the type of the flying object according to a preset classification rule; wherein, the preset classification rule is that if the average temperature is lower than a first temperature threshold and the surface curvature variance is higher than a first curvature threshold, then the flying object is determined to be a bird; if the maximum temperature difference is greater than a second temperature threshold and the surface curvature variance is lower than a second curvature threshold, then the flying object is determined to be a drone. S44, Calculate the confidence level of the classification decision based on the temperature feature and the morphological feature.

[0047] In one possible implementation, for each independent cluster output in step S3, i.e., a point cloud cluster representing a potential flying object, temperature features are extracted from its associated fused point cloud data. The arithmetic mean of the temperature attributes of all data points in the cluster is calculated to reflect the overall thermal radiation level of the target. The difference between the maximum and minimum temperature values ​​of all data points in the cluster is calculated to reflect the uniformity of thermal distribution on or inside the target surface.

[0048] Based on the 3D spatial coordinates of each independent cluster, its morphological features are calculated. Specifically, the 3D convex hull of the clustered point cloud is first calculated, which is the smallest convex polyhedron containing all points, and then the volume of this convex hull is calculated. This feature intuitively reflects the physical dimensions of the target. Based on the clustered point cloud, the curvature at various points on its surface is estimated, and then the variance of these curvature values ​​is calculated. The variance reflects the regularity of the target's surface shape; the larger the variance, the more irregular the surface, such as when a bird flaps its wings; the smaller the variance, the smoother the surface and the more regular the structure, such as the fuselage of a drone.

[0049] The system is equipped with a classification rule base based on prior knowledge to quickly determine the type of flying object based on extracted features. In a specific rule set: if the average temperature is greater than 35 degrees Celsius and the surface curvature variance is greater than 0.15, the target is identified as a bird. This is based on the following: birds' body temperature is typically between 35 and 42 degrees Celsius, with a group average potentially exceeding 35 degrees Celsius; their wing-flapping behavior causes drastic changes in point cloud shape, resulting in a high curvature variance. If the maximum temperature difference is greater than 15 degrees Celsius and the surface curvature variance is less than 0.1, the target is identified as a drone. This is based on the following: drone power components (motors, batteries) generate localized high temperatures, resulting in a significant temperature difference with the environment, leading to a large maximum temperature difference; and their airframe structure is typically relatively regular, with a smooth point cloud surface and a low curvature variance. If the target features do not satisfy any of the rules, or simultaneously satisfy contradictory rules, the system marks it as an unidentified target.

[0050] To assess the reliability of the above rule-based judgments, the system calculates a confidence score for each judgment result. The confidence score calculation takes into account the closeness between the feature value and the judgment threshold.

[0051] If the confidence level of the classification decision is lower than a preset confidence threshold, an adaptive adjustment to the independent clustering is triggered, the adaptive adjustment including: If the temperature gradient within the independent cluster is greater than the preset gradient threshold, then the point cloud data of the independent cluster is clustered a second time according to the temperature gradient to generate a new independent cluster. If the volume of the independent cluster exceeds the preset reasonable volume threshold, the number of clusters determined in S3 is increased, and S33 to S38 are re-executed based on the increased number of clusters to generate a new independent cluster. Based on the new independent clusters obtained, S41 to S43 are re-executed.

[0052] Optionally, the preset confidence threshold can be [70%, 90%], and preferably, the preset confidence threshold can be 80%. Optionally, the preset gradient threshold can be [10℃ / min, 20℃ / min], and preferably, the preset gradient threshold can be 15℃ / min. The classification confidence score calculated in step S44 is compared with a preset confidence threshold. If the confidence score is lower than the preset threshold, the automatic classification result for the current independent cluster is deemed unreliable. The system will automatically trigger a subsequent adaptive adjustment process, aiming to improve the accuracy of subsequent classifications by optimizing the cluster itself. Specifically, the temperature gradient within the independent cluster is calculated. If it is greater than a preset gradient threshold, it indicates that there may be significant differences in heat distribution within the cluster, possibly stemming from different components of a single target, such as the engine and fuselage of a drone, or multiple targets with different temperatures being incorrectly merged. The point cloud data of the cluster is then subjected to secondary clustering based on its temperature attributes. For example, mean clustering or threshold segmentation based on temperature values ​​can be used to divide the original cluster into two or more sub-clusters with more consistent temperature characteristics. These sub-clusters are output as new independent clusters and replace the original independent clusters in subsequent processes.

[0053] The volume of each cluster is compared with a preset reasonable volume threshold. If the volume of an independent cluster exceeds the preset reasonable volume threshold, it indicates that the cluster is physically unreasonable and is very likely composed of multiple neighboring targets. For example, a dense flock of birds may be incorrectly merged into a single cluster by the algorithm in step S3 because their point clouds are spatially adjacent. The number of clusters determined for the current frame in step S33 is increased by an increment. Then, steps S33 to S38 are re-executed with the increased number of clusters. That is, a complete round of dynamic adaptive clustering segmentation is performed again based on the original fused point cloud data and with new parameters. This process aims to separate potentially clustered targets and generate a new set of independent clusters.

[0054] Regardless of which branch or combination of the above methods are used for adjustment, the original processing flow for the original clusters will be interrupted after obtaining new independent clusters. Subsequently, the system will re-execute steps S41 to S43 for these newly generated independent clusters, that is, re-extract the temperature and morphological features of each new cluster and re-determine its category according to the preset classification rules. This loop aims to obtain a more reliable target classification by utilizing the optimized clustering results. If the confidence level after reclassification is still lower than the threshold, it can be adjusted again according to the strategy settings or finally submitted for manual processing.

[0055] This invention addresses the problems of poor adaptability, computational complexity, and the need for large amounts of labeled data in traditional classification models by setting specific threshold classification rules for birds and drones based on prior knowledge, such as temperature thresholds and curvature variance thresholds. It provides a fast, intuitive, and efficient lightweight classification decision mechanism, significantly improving the accuracy and real-time performance of distinguishing between birds and non-biological aircraft, while reducing the system's dependence on computing resources and the complexity of model maintenance. By fusing temperature and morphological features for comprehensive judgment and setting explicit classification logic, it solves the problem of misclassification caused by single feature criteria in complex environments, such as high temperatures against a cold background or regularly shaped organisms. This enhances the robustness and anti-interference ability of classification decisions and reduces the misclassification rate caused by environmental factors or changes in target posture. Furthermore, by addressing the issue of abnormal temperature gradients within clusters when confidence levels are insufficient, it further improves the classification decision mechanism. The system employs a secondary clustering approach based on temperature gradients. This addresses the core issue of mixed feature extraction and classification difficulties caused by significant differences in the thermal characteristics of different components of a single target (such as a drone) or the mismerging of multiple targets with different heat sources. This achieves refined re-segmentation of cluster granularity, effectively improving the purity of feature representation and classification accuracy for targets with internal thermal differences. Furthermore, by increasing the cluster number and re-executing the complete clustering process for independent clusters exceeding a reasonable threshold when confidence is insufficient, the system resolves the fundamental segmentation failure issue caused by the original clustering parameters (cluster number K) potentially being unsuitable for the current dense target scenario, leading to the adhesion of multiple target point clouds and ineffective separation. This endows the system with the ability to dynamically correct segmentation parameters and self-iterate optimization, fundamentally improving the individual separation effect of densely clustered targets (such as flocks of birds), thereby ensuring the accuracy of subsequent target classification and threat assessment.

[0056] S5, based on the type of the flying object, combined with its position and speed, calculate the threat level of the flying object, including: S51, acquire the position information and velocity vector of the flying object, and calculate the minimum spatial distance between the flying object and the center line of the orbit based on the position of the flying object; S52, Based on the velocity vector, calculate the tangential velocity component of the flying object in the direction approaching the orbit; S53, determine the corresponding type weighting coefficient according to the type of the flying object; S54, the threat level of the flying object is calculated by weighted summation based on the minimum spatial distance, the tangential velocity component, and the type weight coefficient; wherein, the weighted summation calculation process is to multiply the reciprocal of the minimum spatial distance term, the tangential velocity component term, and the type weight term after normalization by their respective preset weight coefficients and then sum them.

[0057] In one possible implementation, the location information of successfully classified flying objects is acquired, derived from their spatial coordinates in fused point cloud data. The centerline of the subway track is defined as a series of three-dimensional line segments or smooth curves in a unified world coordinate system. The system calculates the location of the flying object, typically taking the minimum Euclidean distance from its cluster centroid to all line segments on the geometric model of the track centerline, denoted as the minimum spatial distance. This value directly reflects the proximity of the flying object to the safe zone of train operation; the smaller the minimum spatial distance, the higher the potential risk of a spatial collision.

[0058] Obtain the velocity vector of the flying target, which can be obtained from the average velocity of its clustered point cloud. The velocity vector _k and its direction of motion are derived, or estimated through position difference analysis of consecutive frames. To assess the dynamic trend of its approach to the orbit, the system calculates the projection component of this velocity vector in the direction pointing to the nearest point on the orbit centerline, i.e., the tangential velocity component. Specifically, first, the direction vector of the nearest point on the orbit found in step S51 is determined, and then the velocity vector of the object is projected onto this direction. If the projection direction is consistent with the direction pointing to the orbit, the tangential velocity component is positive, indicating that the target is approaching the orbit; otherwise, it is negative or zero. The larger the absolute value of the tangential velocity component, the higher the dynamic risk of the target approaching the orbit.

[0059] Based on the type of flying object determined in step S4, a preset type weight coefficient is assigned to it. This coefficient is set based on the potential hazard level that different types of targets may pose to subway operations. For example, in a weight configuration: the type weight coefficient value is set to 1.0 for targets identified as drones; 0.6 for targets identified as flocks of birds; and 0.3 for targets identified as single birds. For unidentified targets, a medium or high weight value, such as 0.8, can be set to maintain vigilance. This weight reflects the prior knowledge that non-biological intrusions, especially human-controlled drones, are generally considered to have a higher threat level than biological interference.

[0060] To integrate three heterogeneous factors—spatial distance, motion trend, and target type—the system employs a weighted summation model to calculate a unified, dimensionless threat level for an aircraft target. First, each factor is normalized to eliminate dimensional differences and control the numerical range, resulting in distance, velocity, and type terms. The three normalized terms are then weighted and summed. Preset weighting coefficients represent the relative importance of distance, velocity, and type in the overall threat assessment and can be adjusted according to operational safety strategies. The calculated threat level is a scalar value ranging from 0 to a positive number, and its magnitude directly reflects the system's quantitative assessment of the overall threat level of the aircraft target.

[0061] This invention introduces a threat classification concept based on target attributes and prior knowledge by setting differentiated type weight coefficients according to the type of flying object, thereby improving the rationality and refinement of threat assessment results. By employing a weighted summation model to comprehensively calculate the normalized minimum spatial distance reciprocal, tangential velocity component, and type weight, it solves the problems of one-sidedness and inconsistency in threat judgment relying on a single indicator or subjective experience. It provides a quantifiable, repeatable, and parameter-adjustable objective comprehensive threat calculation method, significantly improving the scientificity, systematicity, and interpretability of assessment results. By outputting a quantified comprehensive threat value, it solves the problem that traditional binary early warning or fuzzy descriptions cannot support graded response and accurate decision-making, generating a key quantitative input that can be used to directly drive subsequent automated graded early warning and response strategies, greatly improving the accuracy and automation level of decision support for the entire security monitoring system.

[0062] S6. Determine the risk level based on the threat level of the flying object, and execute the corresponding response strategy based on the risk level.

[0063] The risk level determination based on the threat level of the flying object includes: If the threat level of the flying object is greater than the first threat threshold, it is determined to be of a high-risk level; If the threat level of the flying object is greater than or equal to the second threat threshold and less than or equal to the first threat threshold, it is determined to be of medium risk level; If the threat level of the flying object is less than the second threat threshold, it is determined to be of low risk level.

[0064] The response strategy includes tracing back the trajectory and simulating emergency plans on the preset digital twin platform.

[0065] Preferably, the first threat threshold is 0.8 in a preferred embodiment, and the second threat threshold is 0.5 in a preferred embodiment.

[0066] In one possible implementation, high risk indicates that the target is extremely close to the track, is hurtling toward the track at high speed, or is a high-threat type with an extremely high risk of direct collision or serious interference, requiring immediate intervention at the highest level; medium risk indicates that the target poses a clear potential threat to the track, but the level of urgency has not yet reached its highest level, requiring the activation of early warning and preparation for defensive or decoy measures; low risk indicates that the target currently poses a relatively small immediate threat to the safety of track operations, but its behavior data still needs to be monitored and recorded.

[0067] Based on the determined risk level, pre-set, differentiated response strategies are automatically or assisted. Specifically, for high-risk levels, the target is highlighted in the digital twin platform using the most prominent visual methods, such as a flashing red model or a full-screen warning pop-up. Simultaneously, the highest-level alarm information is automatically generated and pushed to the integrated monitoring system of the subway operation dispatch center via the system interface. Furthermore, based on safety protocol interlocking recommendations, the on-site signaling system can be triggered to enter a special protection mode, such as suggesting trains reduce speed, while simultaneously activating on-site deployed high-intensity sound waves, high-intensity light, signal jammers, or other active deterrent devices to attempt to force the target away from the target area.

[0068] For medium-risk levels, targets are marked with yellow warning icons in the digital twin platform, and warning information is pushed to the handheld terminals or mobile applications of maintenance personnel in the relevant areas. The audible and visual alarms in the area are activated to alert on-site personnel and to notify inspection personnel to go to the relevant areas for verification and monitoring.

[0069] For low-risk targets, the target is displayed calmly in blue or green on the digital twin platform. The system records the target's complete point cloud trajectory, feature data, and low-threat assessment results to the security event database, maintaining routine tracking of the target but without triggering proactive alerts.

[0070] For any target that triggers an alert, operations and maintenance personnel can select the target and time period in the digital twin platform and trigger the trajectory retrospective function with one click. The platform will dynamically reproduce the entire process of the target's appearance, movement, threat assessment, and eventual departure in a 3D scene based on stored historical fused point cloud data and target tracking results. It supports multi-view, slow-motion, and pause viewing, facilitating event review and root cause analysis. The platform has a built-in emergency plan library for various typical intrusion scenarios, such as drone hovering, bird flocks crossing, and kite entanglement. After a real incident occurs, or during routine training, users can call upon relevant plans to simulate and extrapolate within the digital twin environment. The simulation can dynamically display the virtual development process and possible consequences of the event under different response strategies, such as the timing of different expulsion methods, adjustments to train operation plans, and based on historical data or set parameters. This provides powerful visual simulation support for optimizing emergency plans and improving personnel response capabilities.

[0071] This invention addresses the inefficiencies and untimely responses caused by the single, delayed, or overly reliant manual intervention in traditional security systems by pre-setting and automatically triggering differentiated, tiered response strategies for different risk levels. It achieves an automated and collaborative emergency response mechanism that matches the risk level, significantly improving the response speed, appropriateness of measures, and overall prevention and control efficiency in dealing with potential threats, effectively reducing the risk of accident losses due to untimely or inappropriate responses. By deeply integrating early warning information and handling instructions into a digital twin platform for visualization and system linkage, it solves the problems of isolated alarm information, unintuitive presentation, and disconnection from the physical world and operational systems. It achieves real-time mapping and closed-loop between virtual early warning and physical response, greatly enhancing operators' global perception of the threat situation, intuitive decision-making, and smoothness of cross-system collaborative handling. Furthermore, by providing trajectory backtracking functionality in the digital twin platform, it solves the problem of lacking intuitive, continuous, and interactive spatiotemporal data reproduction methods when analyzing and reviewing intrusion events afterward. It provides a powerful and visualized data analysis tool for security incident investigation, responsibility determination, and protection strategy optimization, improving the depth, efficiency, and credibility of event analysis conclusions.

[0072] Example 2 like Figure 2 As shown, the present invention also provides a low-altitude flying object threat identification system based on dynamic density clustering. The system is used to implement any of the low-altitude flying object threat identification methods based on dynamic density clustering described in Embodiment 1. The system includes: The information acquisition module is used to install several monitoring devices at preset intervals in the target area to obtain multimodal monitoring data of the target area. The data fusion module, connected to the information acquisition module, is used to preprocess the multimodal monitoring data and fuse the preprocessed multimodal monitoring data to generate fused point cloud data. The clustering and segmentation module, connected to the data fusion module, is used to perform clustering and segmentation on the fused point cloud data based on an improved clustering algorithm to generate independent clusters; The clustering identification module, connected to the clustering segmentation module, is used to acquire the temperature and morphological features of each independent cluster, and determine the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data. The threat calculation module, connected to the clustering and identification module, is used to calculate the threat level of the flying object based on its type, position, and speed. The risk assessment module, connected to the threat calculation module, is used to determine the risk level based on the threat level of the flying object and to execute corresponding response strategies based on the risk level.

[0073] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for identifying low-altitude flying object threats based on dynamic density clustering, characterized in that, include: S1, Install several monitoring devices at preset intervals in the target area to obtain multimodal monitoring data of the target area; S2, preprocess the multimodal monitoring data, and fuse the preprocessed multimodal monitoring data to generate fused point cloud data; S3, The fused point cloud data is clustered and segmented based on an improved clustering algorithm to generate independent clusters; S4. Obtain the temperature and morphological features of each independent cluster, and determine the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data. S5. Based on the type of the flying object, combined with the position and speed of the flying object, calculate the threat level of the flying object; S6. Determine the risk level based on the threat level of the flying object, and execute the corresponding response strategy based on the risk level; The improved clustering algorithm is used to cluster and segment the fused point cloud data to generate independent clusters, including: S31, Calculate the velocity distribution variance based on the velocity of each point cloud data in the fused point cloud data; S32, Calculate the density value of the point cloud data based on the effective number of points and the detection area; wherein, the effective number of points is the sum of the number of point cloud data with a speed greater than a preset speed threshold; S33, Based on the velocity distribution variance and the density value of the point cloud data, determine the number of clusters for clustering; S34. Based on the cluster centroid positions of historical frames, calculate the minimum cumulative distance sum of each candidate point in the current frame, and select the initial centroid based on the minimum cumulative distance sum; S35, Construct a loss function that integrates spatiotemporal constraints, the loss function including a spatial distance term, a velocity consistency term, and a volume rationality term; S36, with the goal of minimizing the loss function, assigns each point cloud data to the cluster that minimizes its own loss distance, thus completing the clustering assignment; S37, based on the allocation results, update the spatial centroid position, average velocity, and point cloud convex hull volume of each cluster; S38, calculate the current total loss value, and determine whether the difference between the current total loss value and the total loss value of the previous iteration is less than a preset threshold: if yes, stop the iteration and output the independent cluster; if no, return to step S36 to continue the iteration.

2. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 1, characterized in that, The installation of several monitoring devices at preset intervals in the target area includes: S11, At the elevated section of the subway and / or the entrance of the tunnel in the target area, monitoring nodes consisting of millimeter-wave radar and infrared thermal imagers are deployed at preset intervals to form an electronic space curtain covering the target area. S12, radar point cloud data and infrared thermal image data are synchronously collected through the monitoring node as the multimodal monitoring data; S13, based on oblique photography technology, collects real-world images of the surrounding area of ​​the target region and builds them on a pre-set digital twin platform; S14. Based on BIM technology, model the subway vehicles and subway infrastructure to generate a BIM model, and build the BIM model on a preset digital twin platform.

3. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 1, characterized in that, The step of preprocessing the multimodal monitoring data and fusing the preprocessed multimodal monitoring data to generate fused point cloud data includes: S21, preprocess the multimodal monitoring data. The preprocessing includes data cleaning and spatiotemporal alignment. Data cleaning is used to remove invalid and redundant data. Spatiotemporal alignment uses the extended Kalman filter method to unify the monitoring data from different monitoring devices into the same spatiotemporal coordinate system. S22 fuses the preprocessed multimodal monitoring data to generate fused point cloud data containing three-dimensional spatial coordinates, velocity, and temperature information.

4. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 1, characterized in that, The process of acquiring the temperature and morphological features of each independent cluster, and determining the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data, includes: S41, extract the temperature features of each independent cluster, the temperature features including the average temperature of the independent cluster and the maximum temperature difference within the independent cluster; S42, extract the morphological features of each independent cluster, the morphological features including the volume and surface curvature variance calculated based on the point cloud convex hull of the independent cluster; S43, based on the temperature and morphological characteristics, determine the type of the flying object according to a preset classification rule; wherein, the preset classification rule is that if the average temperature is lower than a first temperature threshold and the surface curvature variance is higher than a first curvature threshold, then the flying object is determined to be a bird; if the maximum temperature difference is greater than a second temperature threshold and the surface curvature variance is lower than a second curvature threshold, then the flying object is determined to be a drone. S44, Calculate the confidence level of the classification decision based on the temperature feature and the morphological feature.

5. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 4, characterized in that, If the confidence level of the classification decision is lower than a preset confidence threshold, an adaptive adjustment to the independent clustering is triggered, the adaptive adjustment including: If the temperature gradient within the independent cluster is greater than the preset gradient threshold, then the point cloud data of the independent cluster is clustered a second time according to the temperature gradient to generate a new independent cluster. If the volume of the independent cluster exceeds the preset reasonable volume threshold, the number of clusters determined in S3 is increased, and S33 to S38 are re-executed based on the increased number of clusters to generate a new independent cluster. Based on the new independent clusters obtained, S41 to S43 are re-executed.

6. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 1, characterized in that, The calculation of the threat level of a flying object based on its type, position, and speed includes: S51, acquire the position information and velocity vector of the flying object, and calculate the minimum spatial distance between the flying object and the center line of the orbit based on the position of the flying object; S52, Based on the velocity vector, calculate the tangential velocity component of the flying object in the direction approaching the orbit; S53, determine the corresponding type weighting coefficient according to the type of the flying object; S54, the threat level of the flying object is calculated by weighted summation based on the minimum spatial distance, the tangential velocity component, and the type weight coefficient; wherein, the weighted summation calculation process is to multiply the reciprocal of the minimum spatial distance term, the tangential velocity component term, and the type weight term after normalization by their respective preset weight coefficients and then sum them.

7. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 1, characterized in that, The risk level determination based on the threat level of the flying object includes: If the threat level of the flying object is greater than the first threat threshold, it is determined to be of a high-risk level; If the threat level of the flying object is greater than or equal to the second threat threshold and less than or equal to the first threat threshold, it is determined to be of medium risk level; If the threat level of the flying object is less than the second threat threshold, it is determined to be of low risk level.

8. The method for identifying low-altitude flying object threats based on dynamic density clustering according to claim 1, characterized in that, The response strategy includes tracing back the trajectory and simulating emergency plans on the preset digital twin platform.

9. A low-altitude flying object threat identification system based on dynamic density clustering, the system being used to implement the low-altitude flying object threat identification method based on dynamic density clustering as described in any one of claims 1 to 8, the system comprising: The information acquisition module is used to install several monitoring devices at preset intervals in the target area to obtain multimodal monitoring data of the target area. The data fusion module, connected to the information acquisition module, is used to preprocess the multimodal monitoring data and fuse the preprocessed multimodal monitoring data to generate fused point cloud data. The clustering and segmentation module, connected to the data fusion module, is used to perform clustering and segmentation on the fused point cloud data based on an improved clustering algorithm to generate independent clusters; The clustering identification module, connected to the clustering segmentation module, is used to acquire the temperature and morphological features of each independent cluster, and determine the type of flying object based on the temperature and morphological features of the independent clusters and the corresponding fused point cloud data. The threat calculation module, connected to the clustering and identification module, is used to calculate the threat level of the flying object based on its type, position, and speed. The risk assessment module, connected to the threat calculation module, is used to determine the risk level based on the threat level of the flying object and to execute corresponding response strategies based on the risk level.

Images